Self-Optimized One-Class Classification Using Sum of Ranking Differences Combined with a Receiver Operator...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• One-class classification (class modeling) is crucial for tasks like food authentication and medical diagnosis.
• Selecting and optimizing one-class classifiers, alongside choosing appropriate spectroscopic data preprocessing and spectral regions, presents significant challenges.
• Existing methods often require extensive optimization for each classifier and dataset.

Purpose of the Study:

• To develop a hybrid fusion process capable of combining multiple, non-optimized one-class classifiers.
• To address the challenges of classifier selection, parameter tuning, and data preprocessing in spectroscopic analysis.
• To demonstrate a flexible and reliable approach for class modeling using spectroscopic data.

Main Methods:

• A hybrid fusion process combining non-optimized classifiers across multiple instruments, preprocessing methods, and measurements.
• Utilizing a window of tuning parameters for each classifier instead of full optimization.
• Applying the sum of ranking differences (SRD) method for flexible fusion of assessment values.
• Automatically optimizing the SRD ranking value (threshold) using a receiver operator characteristic (ROC) curve.

Main Results:

• The hybrid fusion approach demonstrated reliable classification performance on two distinct analytical datasets.
• The first dataset involved beer authentication using five instruments (NIR, MIR, UV-Vis, TGA), with three fusion protocols tested.
• The second dataset analyzed MIR spectra of strawberry puree for authenticity, achieving reliable accuracy, sensitivity, and specificity.

Conclusions:

• Fusing non-optimized classifiers offers a reliable strategy for one-class classification in analytical chemistry.
• The proposed hybrid fusion method, utilizing SRD and ROC curve optimization, effectively handles multi-instrument spectroscopic data.
• This approach provides a robust solution for class modeling problems, improving classification reliability without extensive individual classifier optimization.